Energy-Based Reasoners: 385x more efficient inference
|
Energy-Based Reasoners (EBRs) introduce a novel approach to multi-step reasoning by learning an energy function that guides iterative refinement of latent states, providing explicit signals for when to stop and which solution to select. Across algorithmic reasoning benchmarks like Sudoku-Extreme, EBRs achieve state-of-the-art accuracy (99.9%) while using 345× less inference compute than comparable models, demonstrating that energy-based models are highly efficient for reasoning tasks.